Repo Agent Kit RSS
All field notes
Dataset · field note5 min read

A pinned snapshot found root AGENTS.md in 39 of GitHub’s top 100 repositories

What a reproducible September 2026 snapshot reveals about root AGENTS.md adoption, file shape, testing guidance, scope rules, and completion criteria.

Practical takeaway

Adoption is already visible among highly starred repositories, but a root file alone is not evidence that its instructions are complete or operational.

A dated ranking needs pinned evidence

A top-100 claim becomes stale as soon as repository stars or default branches move. The snapshot therefore records the first 100 public repositories returned by GitHub search sorted by stars on September 1, 2026, then resolves every default branch to an exact commit before checking for the file.

Thirty-nine repositories contained an exact, case-sensitive AGENTS.md at the root of that pinned commit. Sixty-one did not. The result deliberately ignores nested AGENTS.md files when the root file is absent, so it measures one discoverable convention rather than every instruction file in each repository.

Presence and instruction shape answer different questions

The 39 root files had a median length of 112 lines. Twenty-nine matched a bounded testing, validation, or quality-guidance signal; 32 matched scope, boundary, or explicit-restriction language; and only 10 matched the completion-guidance signal used by the collector.

Those matches are structural observations, not quality scores. A keyword can be vague, incorrect, or obsolete. Conversely, useful instructions can use language the bounded expressions do not match. The separate readiness benchmark exists because adoption and operational quality should not be collapsed into one number.

  • 39 of 100 had an exact root AGENTS.md.
  • The median adopted file contained 112 lines.
  • Testing and scope signals appeared more often than completion guidance.

Make the result reusable, not just quotable

The public dataset includes all 100 repository rows, exact commit links, file evidence for positive rows, repository metadata, line and heading counts, command-shaped line counts, and the bounded signals. It does not reproduce the instruction-file contents.

JSON and CSV downloads make the snapshot independently inspectable. The collector script is kept with the product source so a later snapshot can use the same inclusion, pinning, and structural-analysis rules instead of silently changing the method.